Your Observations? The first kernel seems to be detecting edges in the image. Diagonal edges give a stronger response. The kernel supports this guess because it supports a stronger magnitude on the diagonals, but if pixels of similar scale are on all 4 corners, it sums to ~0 The second kernel seems to be finding local maxima, including edges, corners. This is because the 4 multiply in the center will overcome the 4 other pixels that contribute if one of those is less than the others. If one or more of the -1 pixels is a greater magnitude than the center, however, it isn't necessarily enough to counter the center, so it seems to effect non-local maxima as well.